fix(model_metadata): cache detect_local_server_type result for process lifetime

[uzaylisak]

Problem

Every 5 minutes fetch_endpoint_model_metadata() is called (TTL = _ENDPOINT_MODEL_CACHE_TTL = 300). Each time it calls detect_local_server_type(), which blindly re-runs the full probe waterfall:

1. GET /api/v1/models → 404 (LM Studio check)
2. GET /api/tags → 404 (Ollama check)
3. GET /v1/props → 404 (llama.cpp check)
4. GET /version → 200 (vLLM — found)

On a vllm backend the first three always 404, polluting server logs every 5 minutes. Reported in #29971.

Fix

Add _endpoint_probe_path_cache: Dict[str, str] = {} at module level. detect_local_server_type() checks this cache on entry; on a hit it returns immediately without making any HTTP requests. On the first call the waterfall runs as before, and if a server type is identified it is stored for the lifetime of the process.

The probe stages also use if result is None guards so each step is skipped once an earlier one succeeds — reducing requests even on the very first call.

Impact

• vllm (and any server that is not LM Studio or Ollama): only 1 probe round-trip ever, vs. 4 per 5-minute refresh cycle previously.
• LM Studio, Ollama, llama.cpp: one extra dict lookup per call; behaviour is identical.
• Cache is intentionally process-scoped (not TTL-based) because a server type does not change during a session.

Closes #29971.

[alt-glitch]

Note: closed PR #29986 was the first attempt at this same fix.

[kshitijk4poor]

Thanks @uzaylisak — your fix landed! The probe-cache cluster has now landed on main via PR #61368 (merge commit f556edc), which salvaged this cluster of PRs onto current main with structured review, live smoke tests, and full test gates.

Your commit was cherry-picked with -x so your authorship is preserved on main (commit a49c48927, fix(model_metadata): cache detect_local_server_type result for process lifetime). During review the cache was tightened from process-lifetime to a bounded 1h TTL so a backend swap on the same port is eventually re-detected, with regression tests covering both the cache hit and TTL-expiry re-detection.

Closing since this is now merged with your authorship intact.